Privacy-provisioned-permission-aware RAG in cloud using identity and access management, trust and large-language models (PPPARITL)

Abstract Reasoning-based decision making in LLMs often suffer from hallucinations and tend to generate outdated information since they lack domain-specific knowledge. Retrieval-Augmented Generation (RAG) models enhance the quality of contextual information by anchoring the information retrieval to facts illustrated in the Knowledge Sources (KS). However, since business enterprises rely immensely on remote cloud to handle open-ended queries from service users and process data outsourcing, significant authorization risks and privacy concerns persist. To address these concerns, this work proposes a Privacy-Provisioned-Permission-Aware RAG that integrates Identity and Access Management and Trust to facilitate decision-making in Large-Language Models (PPPARITL). Unlike traditional RAG models, the PPPARITL framework protects document and query embeddings at rest using CKKS homomorphic encryption; embeddings are decrypted only by an administrator-held secret key, and only for candidates already authorized under the RRAC gate, before similarity ranking is performed. Each document (structured or unstructured) in the KS is protected by enforcing constructive alignment between the user-role mapping in the Access Control List (ACL). Further, the work also introduces a sanitization layer to handle Denial-of-Service (DoS) attack and revoke malicious query patterns. The framework works in 2 prominent phases- In phase 1, each user query is validated with the pre-defined Identity and Access Management (IAM) policy rules configured by the administrator as the non-compensatory Role-Resource-Action-Condition (RRAC) gate in the ACL. To safeguard the information leakage, a trust score is computed in the later phase for each service request analysing contextual, behavioral, and historical service user interactions governing Permission-Aware-Retrieval. It is this scenario-based trust assignment that facilitates LLM to ensure privileged access and provide explainable and intelligent reasoning. In addition, the PPPARITL framework continuously monitors malicious queries to mitigate internal threats and revoke users if needed. The experiments show that the proposed work outperforms the baseline access control approaches, exhibiting an increased authorization accuracy of 95.84% on the synthetic enterprise dataset and a substantial decrease in the False Access Rate (FAR). These simulation results maintain an effective balance between trust-oriented retrieval and dependable access control systems for cloud ecosystems.